{ "id": 99580, "title": "[HELP]loss is decreasing, but still large. I do not know why?", "url": "/competitions/aptos2019-blindness-detection/discussion/99580", "author_name": "DeepInvolution", "post_date": "2019-07-12T10:13:26.637000", "votes": 0, "comment_count": 0, "views": 0, "content": "

Hi there,

\n\n

I built a CV pipeline to train. Special thanks to Abhishek's kernel because very much the same model of Abihishek's. However, the predicted labels of valid dataset are very different from the target (before optimizer of quadratic weighted kappa). The distribution of regression result is as follows:\n\"\"

\n\n

Also, loss is decreasing as show below (trained for 3 epochs but it is decreasing for 20 epochs), but it is much larger, around 1.68, than the loss of Abihishek's kernel which is under 0.6. I ran out of ideas to find out the reason and almost went mad.\n\"\"

\n\n

Anyone kindly give some ideas or tips? Thanks in advance!

\n\n

Model: resnet101

", "messages": [ { "id": 573469, "postDate": "2019-07-12T10:13:26.637Z", "content": "

Hi there,

\n\n

I built a CV pipeline to train. Special thanks to Abhishek's kernel because very much the same model of Abihishek's. However, the predicted labels of valid dataset are very different from the target (before optimizer of quadratic weighted kappa). The distribution of regression result is as follows:\n\"\"

\n\n

Also, loss is decreasing as show below (trained for 3 epochs but it is decreasing for 20 epochs), but it is much larger, around 1.68, than the loss of Abihishek's kernel which is under 0.6. I ran out of ideas to find out the reason and almost went mad.\n\"\"

\n\n

Anyone kindly give some ideas or tips? Thanks in advance!

\n\n

Model: resnet101

", "rawMarkdown": "Hi there,\n\nI built a CV pipeline to train. Special thanks to [Abhishek's kernel](https://www.kaggle.com/abhishek/very-simple-pytorch-training-0-59) because very much the same model of Abihishek's. However, the predicted labels of valid dataset are very different from the target (before optimizer of quadratic weighted kappa). The distribution of regression result is as follows:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F727580%2Fc6a4bbd877755bb7dc16fb477dae674b%2Findex.png?generation=1562926053100251&alt=media)\n\nAlso, loss is decreasing as show below (trained for 3 epochs but it is decreasing for 20 epochs), but it is much larger, around 1.68, than the loss of Abihishek's kernel which is under 0.6. I ran out of ideas to find out the reason and almost went mad.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F727580%2F2a792dc8faba73de51c74f3e1bb342ee%2Findex1.png?generation=1562926242621833&alt=media)\n\nAnyone kindly give some ideas or tips? Thanks in advance!\n\nModel: resnet101" } ], "comments": [], "raw_markdown_by_id": { "573469": "Hi there,\n\nI built a CV pipeline to train. Special thanks to [Abhishek's kernel](https://www.kaggle.com/abhishek/very-simple-pytorch-training-0-59) because very much the same model of Abihishek's. However, the predicted labels of valid dataset are very different from the target (before optimizer of quadratic weighted kappa). The distribution of regression result is as follows:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F727580%2Fc6a4bbd877755bb7dc16fb477dae674b%2Findex.png?generation=1562926053100251&alt=media)\n\nAlso, loss is decreasing as show below (trained for 3 epochs but it is decreasing for 20 epochs), but it is much larger, around 1.68, than the loss of Abihishek's kernel which is under 0.6. I ran out of ideas to find out the reason and almost went mad.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F727580%2F2a792dc8faba73de51c74f3e1bb342ee%2Findex1.png?generation=1562926242621833&alt=media)\n\nAnyone kindly give some ideas or tips? Thanks in advance!\n\nModel: resnet101" } }